Photonic Neural Activation Function Based on High-Speed Electro-Absorption Modulated Laser

Abstract

Integrated optics hold great potential to accelerate deep learning tasks with high clock rates, parallelism and low-loss data transmission. Silicon photonic integrated circuits can perform large-scale and low-power-consuming optical linear operations by using weighting mechanism through linear optics. However, on-chip light attenuation and nonlinear activation functions are still huge challenges for large-scale optical neural networks. Here, we demonstrate a high-speed electro-absorption modulator (EAM) monolithically integrated with a distributed feedback (DFB) laser that can deliver high output lasing power for larger scale expansion while acting as a nonlinear activation function unit. With the use of the obtained nonlinear activation function, a convolutional neural network (CNN) is simulated to perform a handwritten digit classification benchmark task with high accuracy. Thanks to its compactness, high response speed and laser integration, the demonstrated nonlinear unit has the potential to be used in heterogeneously integrated large-scale photonic neural networks.